All forms of Artificial Intelligence are not the same – there are different approaches to how the systems map their knowledge. A distinction is made primarily between two methods: neuronal networks and symbolic AI.

Knowledge represented by symbols

Conventional AI is mainly about logical analysis and planning of tasks. Symbolic, or rule-based, AI is the original method developed back in the 1950s. It attempts to simulate human intelligence by processing abstract symbols and with the aid of formal logic. This means that facts, events or actions are represented by concrete and unambiguous symbols. Based on these symbols, mathematical operations can be defined, such as the programming paradigm “if X, then Y, otherwise Z”. The knowledge – that is to say, the sum of all symbols – is stored in large databases against which the inputs are cross-checked. The databases must be “fed” in advance by humans. Classic applications of -symbolic AI include, for example, text processing and voice recognition. Probably the most famous example of -symbolic AI is DeepBlue, IBM’s chess computer which beat then world champion Garry Kasparov using symbolic AI in 1997.

As computer performance increases steadily, symbolic AI is able to solve ever more complex problems. It works on the basis of fixed rules, however. For a machine to operate beyond tightly constrained bounds, it needs much more flexible AI capable of handling uncertainty and processing new experiences.

Advancing knowledge about neurons autonomously

That flexibility is offered by artificial neuronal networks, which are currently the focus of research activity. They simulate the functionality of the human brain. Like in nature, artificial neuronal networks are made up of nodes, known as neurons, or also units. They receive information from their environment or from other neurons and relay it in modified form to other units or back to the environment (as output). There are three different kinds of unit:

Input units receive various kinds of information from the outside world. This may be measurement data or image information, for example. The data, such as a photo of an animal, is analysed across multiple layers by hidden units. At the end of the process, output units present the result to the outside world: “The photo shows a dog.” The analysis is based on the edge by which the individual neurons are interconnected. The strength of the connection between two neurons is expressed by a weight. The greater the weight, the more one unit influences another. Thus the knowledge of a neuronal network is stored in its weights. Learning normally occurs by a change in weight; how and when a weight changes is defined in learning rules. So before a neuronal network can be used in -practice, it must first be taught those learning rules. Then neuronal networks are able to apply their learning algorithm to learn independently and grow autonomously. That is what makes neuronal AI a highly dynamic, adaptable system capable of mastering challenges at which symbolic AI fails.

Cognitive processes as the basis of a new AI

Another new form of Artificial Intelligence has been developed by computer scientists at the University of -Tübingen: their “Brain Control” computer program simulates a 2D world and virtual figures – or agents – that act, cooperate and learn autonomously within it. The aim of the simulation is to translate state-of-the-art cognitive science theories into a model and research new variants of AI. Brain Control has not made use of neuronal networks to date, but nor does it adhere to the conventional AI paradigm. The core theoretical idea underlying the program originates from a cognitive psychology theory, according to which cognitive processes are essentially predictable, and based on so-called events. According to the theory, events – such as a movement to grip a pen, and sequences of events such as packing up at the end of the working day – form the building blocks of cognition, by which interactions, and sequences of interactions, with the world are selected and controlled in a goal-oriented way. This hypothesis is mirrored by Brain Control: the figures plan and decide by simulating events and their sequencing, and are thus able to carry out quite complex sequences of actions. In this way, the virtual figures can even act collaboratively. First, one figure places another figure on a platform so that the second figure can clear the way, then both of them are able to advance. The modelling of cognitive systems such as in Brain Control is still an ambitious undertaking. But its aim is to deliver improved AI over the long term.

The post Neuronal networks simulating the brain appeared first on Future Markets Magazine.

Many functions of intelligent information systems are based on methods of pattern recognition: support for diagnoses in the area of medicine, voice recognition with assistance systems and translation tools, object detection in camera images and videos or also forecasting of stock prices. All of these applications involve identifying patterns – or rules – in large volumes of data. It is immaterial whether this data relates to information stored in a database or to pixels in an image or the operating data of a machine. Such identification of patterns was either not possible at all with classic computer systems or required lengthy calculation times of up to several days.

Classifying data in seconds

Developments in the area of neural networks and machine learning have led to the emergence of solutions today in which even complex input data can be matched and classified within minutes or even seconds with trained features. A distinction is made here between two fundamental methods: supervised and unsupervised classification.

With supervised classification of input data in pattern recognition, the system is “fed” training data, with the data with the correct result being labelled accordingly. The correct response must therefore be available during the training phase and the pattern recognition algorithm has to fill the gap between input and output. This form of supervised pattern recognition is used with machine vision for object detection or for facial recognition for example.

In the case of unsupervised learning, the training data is not labelled, which means that the possible results are not known. The pattern-recognition algorithm therefore cannot be trained by providing it with the results it is to arrive at. Algorithms are used more so, which explore the structure of the data and derive meaningful information from it. To stay with the example of machine vision: the techniques of unsupervised pattern recognition are used for object detection, among other things. Unsupervised methods are essentially also used for data mining, thus for detecting contents in large data volumes based on visibly emerging structures.

Finding structures in big data

A number of different methods are in turn used in this type of big data analysis. One such example is association pattern analysis. A set of training data is searched through in this case for combinations of individual facts or events, which occur significantly often or significantly rarely together in the data. Another example in this context is what is known as sequential pattern mining. A set of training data is searched through to identify time-ordered sequences that occur conspicuously often or rarely in succession in the data. The result of the different mining methods is a collection of patterns or rules, which can be applied to future data sets to discover whether one or more rules occur in these data sets. The rules can be integrated in operative software programs in order to develop early warning concepts, for example, or to predict when maintenance is due.

The post Pattern recognition trough AI appeared first on Future Markets Magazine.

Artificial Intelligence is currently a real buzzword, even – or especially – beyond the high-tech industries. But all AI is not the same. Experts classify Artificial Intelligence in three different forms:

1. Weak Artificial Intelligence
2. Strong Artificial Intelligence
3. Artificial Supertintelligence

First, there is weak AI (also known as narrow AI). This type of Artificial Intelligence is only able to perform a single, specific, clearly defined task. It employs mathematical and computer science methodology; its approach to problem-solving is not varied, and it gains no deeper understanding of the problem at hand. This means weak AI is quite capable of performing tasks better than a human, but it can’t be used to solve any problems other than the original defined task. Familiar examples of weak AI are Siri, some robots used in industrial manufacturing, and autonomous vehicles.

Same intellectual abilities as a human

When most people think of AI, they think of machines that are more intelligent than humans and can do anything a human can. That type of Artificial Intelligence is termed strong AI (also known as general AI). Machines with strong AI can attain the same intellectual abilities as a human or even surpass them. AI of this type is able to solve more than just one specific problem, act on its own initiative, plan, learn, and communicate in natural language. It has not yet been possible to develop strong AI to date, however. But scientists do consider it a realistic prospect within the next 20 to 40 years.

When machines become more intelligent than their creators

The third category of AI is Artificial Superintelligence (ASI). It will have been attained when machines surpass the intelligence of the most intelligent minds in human history. The fear is that a stage will then have been reached where humans are no longer the dominant species. Super-intelligent machines would be capable of building still better machines; AI would advance explosively, leaving human intelligence far behind. This scenario has become known as the “technological singularity”. Although it was first coined back in the 1960s, the term became particularly popular in 1998 through a book titled “The Singularity is Near” by Raymond Kurzweil. In his book, Kurzweil predicts that the technological singularity will be reached in the year 2045. That is when he estimates that the computing power of AI will have surpassed the intelligence of all of humanity by a factor of a billion. Different time-scales have repeatedly been mooted since Kurzweil’s book appeared, however. There is currently no sign of Artificial Superintelligence on the horizon. Yet we should not forget that computing power is advancing at a rapid rate; the performance of computer chips has doubled every 18 months in the past. So computers are advancing much more rapidly than human consciousness. Whereas humans have taken millennia to develop, computers have done so in less than 100 years. So, it seems only a matter of time before Artificial Superintelligence arrives…

The post Forms of Artificial Intelligence appeared first on Future Markets Magazine.

Many experts regard this area as currently having the greatest potential within AI.

Deep learning currently dominates ­learning methods

Deep learning is a method of machine learning: many experts regard this area as currently having the greatest potential within AI. Deep learning uses complex neural networks to learn autonomously how something can be classified. The system records large volumes of known information – for example pictures or sounds – in a database and compares it with unknown data.

The procedure eliminates many work steps involved in classic machine learning. That’s because the training effort is significantly less: the “trainer” simply has to present the neural network with data such as pictures – the system discovers for itself how the objects shown in the pictures are to be classified. The human has to unambiguously indicate whether the object whose recognition is to be learned can be seen in the picture (therefore, for example, whether or not a pedestrian is shown in the picture). The ­deep-learning program uses the information from the training data in ­order to define typical features of a pedestrian and generate a prediction model from this. The system works down deeper into the neural network level by level – hence the name deep learning. The nodes at the first level, for example, only register the brightness values of the image pixels. The next level recognises that some of the pixels form lines. The third differentiates between horizontal and vertical lines. This iterative process continues until the system recognises legs, arms and faces and has learned how a person in the picture should be classified.

This learning process requires significant computing power, however, and therefore places increased demands on the processor technology. Researchers and ­manufacturers are consequently working intensively on developing ­special ­AI chips that can perform even more computing processes faster.

Simply a case of sharing acquired ­knowledge

At the same time, thought is being given as to how the knowledge that a system has elaborately acquired can be made available to other systems, too. This would then mean, for example, that not every autonomous vehicle would have to learn for itself what a pedestrian looks like, rather it could draw on the experience of vehicles that have been on the road for longer. The Khronos Group, an open consortium of leading hardware and software companies presented an exchange format for neural networks at the end of 2017. The Neural Network Exchange Format 1.0 allows scientists and engineers to transfer existing trained networks from the training platform to a host of other systems – in other words, in the same way as with PDF format in text processing.

The post Deep Learning in Artificial Intelligence appeared first on Future Markets Magazine.

Congested streets, rising emission levels and the lack of parking all combine to make urban logistics an increasingly greater challenge. Powered by e-commerce, the package market is growing by seven to ten per cent annually in mature markets such as the United States or -Germany. This will see the volume double in Germany by 2025, with around five billion packages mailed each year. “While -deliveries to consumers have previously made up about 40 per cent, more than half of all packages are now delivered to private households. Timely delivery in ever greater -demand,” says Jürgen Schröder, a McKinsey -Senior Partner and expert in logistics and postal services. “New technologies like autonomous driving and drone delivery still need to be developed further. They present opportunities to reduce costs and simplify delivery. We expect that by 2025, it will be possible to deliver around 80 per cent of packages by automated means.”

Package-carrying drones, as Amazon put forward for the first time in 2013, were initially laughed off as a crazy idea. Today, a large number of companies are experimenting with delivery by drone. One example is Mercedes-Benz with its Vans & Drones concept, in which the package is not directly delivered to the customer via a drone, but in a commercial vehicle. In the summer of 2017, the company carried out autonomous drone missions for the first time in an urban environment in Zurich. In the course of the pilot project, -customers could order selected products on Swiss online marketplace siroop. These were a -maximum of two -kilograms in weight and suitable for transport by drone. The range of products included coffee and -electronics. The -customers received their goods the same day. The retailer loaded the drones immediately after receiving the -order on its own premises. After this, they flew to one of two Mercedes vans used in the project, which featured an -integrated drone landing platform. The vans stopped at one of four -pre–determined “rendezvous points” in the Zurich metropolitan area. At these points, the mail carriers received the products and delivered them to the customers, while the drone returned to the retailer. Overall, some 100 flights were -successfully completed without any incidents across the urban area. “We believe that drone-based logistics networks will fundamentally change the way we access products on a daily basis,” says Andreas Raptopoulos, Founder and CEO of Matternet, the manufacturer of the drones used in the test.

Reliably dodging obstacles thanks to Artificial Intelligence

An essential element of such applications are drones that can fly safely between buildings or in a dense street network, where cyclists and pedestrians can suddenly cross their path. Researchers at the University of Zurich and the NCCR Robotics research centre developed an intelligent solution for this purpose. Instead of relying on sophisticated -sensor systems, the drone developed by the Swiss -researchers uses a standard smartphone camera and a very -powerful AI -algorithm called DroNet. “DroNet recognises static and dynamic obstacles and can slow down to avoid crashing into them. With this algorithm, we have taken a step forward towards integrating autonomously navigating drones into our everyday life,” explains Davide Scaramuzza, Professor for Robotics and Perception at the University of Zurich. For each input image, the algorithm generates two outputs: one for navigation to fly around obstacles and one for the likelihood of collisions to detect dangerous situations and make it possible to respond. In order to gain enough data to train the algorithm, information was collected from cars and bicycles that were travelling in urban environments in accordance with the traffic rules. By imitating them, the drone -automatically learned to respect the safety rules, for example “How do we follow the street without crossing into the oncoming lane” or “How do we stop when obstacles like pedestrians, construction works or other vehicles block the way?”. Having been trained in this way, the drone is not only capable of navigating roads, but also of finding its way around in completely different environments than those it was ever trained for – such as multi-storey car parks or office corridors.

Drones controlled by Artificial Intelligence are winning the race

Just how sophisticated drones controlled by Artificial Intelli-gence are today was demonstrated in a race organised by NASA’s Jet Propulsion Laboratory (JPL), when world-class drone pilot Ken Loo took on Artificial Intelligence in a timed trial. “We pitted our algorithms against a human, who flies a lot more by feel,” said Rob Reid of JPL, the project’s task manager. Compared to Loo, the drones flew more cautiously but consistently. The drones needed around 3 seconds longer for the course, but kept their lap times constant at a speed of up to 64 kilometres per hour, while the human pilot varied greatly and was already exhausted after a few laps.

(Picture Credit: Daimler AG)

The post Drones controlled by AI appeared first on Future Markets Magazine.

We are teaching cars to negotiate road traffic autonomously,” reports Dr Volkmar Denner, Chairman of the Board of Bosch. “Automated driving makes the roads safer. Artificial Intelligence is the key. Cars are becoming smart,” he asserts. As part of those efforts, the company is currently developing an on-board vehicle computer featuring AI. It will enable self-driving cars to find their way around even complex environments, including traffic scenarios that are new to it.

Transferring knowledge by updates

The on-board AI computer knows what pedestrians and cyclists look like. In addition to this so-called object recognition, AI also helps self-driving vehicles to assess the situation around them. They know, for example, that an indicating car is more likely to be changing lane than one that is not indicating. This means self-driving cars with AI can detect and assess complex traffic scenarios, such as a vehicle ahead turning, and apply the information to adapt their own driving. The computer stores the knowledge gathered while driving in artificial neuronal networks. Experts then check the accuracy of the knowledge in the laboratory. Following further testing on the road, the artificially created knowledge structures can be downloaded to any number of other on-board AI computers by means of updates.

Assistants recognising speech, gestures and faces

Bosch is also intending to collaborate with US technology company Nvidia on the design of the central vehicle computer. Nvidia will supply Bosch with a chip holding the algorithms for the vehicle’s movement created through machine learning. As Nvidia founder Jensen Huang points out, on-board AI in cars will not only be used for automated driving: “In just a few years, every new vehicle will have AI assistants for speech, gesture and facial recognition, or augmented reality.” In fact, the chip manufacturer has also been working with Volkswagen on the development of an intelligent driving assistant for the electric microvan I.D.Buzz. It will process sensor data from both inside the car and from its surrounding environment to control comfort and assistance systems. These systems will be able to accumulate new capabilities in the course of further developments in autonomous driving. Thanks to deep learning, the car of the future will learn to assess situations precisely and analyse the behaviour of other road users.

3D recognition using 2D cameras

Key to automated driving is creating the most exact map possible of the surrounding environment. The latest camera systems are also using AI to do that. A project team at Audi Electronics Venture, for example, has developed a mono camera which uses AI to generate a high-precision three-dimensional model of the surroundings. The sensor is a standard, commercially available front-end -camera. It captures the area in front of the car to an angle of about 120 degrees, taking 15 frames per second at a 1.3 megapixel resolution. The images are then processed in a neuronal network. That is also where the so-called semantic segmentation takes place. In this process, each pixel is assigned one of 13 object classes. As a result, the system is able to recognise and distinguish other cars, trucks, buildings, road markings, people and traffic signs. The system also uses neuronal networks to gather distance information. This is visualised by so-called ISO lines – virtual delimiters which define a constant distance. This combination of semantic segmentation and depth perception creates a precise 3D model of the real environment. The neuronal network is pre-trained through so-called unsupervised learning, having been fed with large numbers of videos of road scenarios captured by a stereo camera. The network subsequently learned autonomously to understand the rules by which it generates 3D data from the mono camera’s images.

Mitsubishi Electric has also developed a camera system that uses AI. It will warn drivers of future mirrorless vehicles of potential hazards and help avoid accidents, especially when changing lane. The system uses a new computing model for visual recognition that copies human vision. It does not capture a detailed view of the scene as a whole, but instead focuses rapidly on specific areas of interest within the field of vision. The relatively simple visual recognition algorithms used by the AI conserve the system resources of the on-board computer. The system is nevertheless able to distinguish between object types such as pedestrians, cars and motorcycles. Compared to conventional camera-based systems, the technology will be able to significantly extend the maximum object recognition range from the current approximately 30 metres to 100 metres. It will also be able to improve the accuracy of object recognition from 14 percent to 81 per cent.

AI is becoming a competitive factor

As intelligent assistance systems are being implemented ever more frequently, AI is becoming a key competitive factor for car manufacturers. That is true with regard to the use of AI for autonomous driving as well as in the development of state-of-the-art mobility concepts based on AI. According to McKinsey, almost 70 per cent of customers are already willing to switch manufacturer today in order to gain better assisted and autonomous driving features. The advice from Dominik Wee, Partner at McKinsey’s Munich office, is: “Premium manufacturers, in particular, need to demonstrate to their highly demanding customers that they are technology leaders in AI-based applications as in other areas – for example in voice-based interaction with the vehicle, or in finding a parking space.”

(Picture Credit: Volkswagen AG)

The post Autonomous driving thanks to AI appeared first on Future Markets Magazine.

The field of Artificial Intelligence is developing rapidly, and promises to help address some of the biggest challenges we face as a society,” says Kate Crawford, Co-Founder of the AI Now Institute: “But we urgently need more research into the real-world implications of the adoption of AI in our most sensitive social institutions. People are already being affected by these systems, be it while at school, looking for a job, reading news online, or interacting with the courts.” It is for this very reason that the AI Now Institute was launched in late 2017 at New York University. This is the first university research institute dedicated to the social impact of Artificial Intelligence. To this end, it wants to expand the scope of AI research to include experts from fields such as law, healthcare, occupational and social sciences. According to Meredith Whittaker, another Co-Founder of AI Now, “AI will require a much wider range of expertise than simply technical training. Just as you wouldn’t trust a judge to build a deep neural network, we should stop assuming that an engineering degree is sufficient to make complex decisions in domains like criminal justice. We need experts at the table from fields like law, healthcare, education, economics, and beyond.”

Safe and just AI requires a much broader spectrum of expertise than mere technical know-how.

AI Systems with Prejudices are a Reality

“We’re at a major inflection point in the development and implementation of AI systems,” Kate Crawford states. “If not managed properly, these systems could also have far-reaching social consequences that may be hard to foresee and difficult to reverse. We simply can’t afford to wait and see how AI will affect different populations.” With this in mind, the AI Now Institute is looking to develop methods to measure and understand the impacts of AI on society.

It is already apparent today that unsophisticated or biased AI systems are very real and have consequences – as shown, in one instance, by a team of journalists and technicians at Propublica, a non-profit newsdesk for investigative journalism. They tested an algorithm which is used by courts and law enforcement agencies in the United States to predict repeat offending among criminals. They found that it was measurably biased against African Americans. Such prejudice-laden decisions come about when the data that the AI is based on and works with is not neutral. If it includes social disparities, for instance, the evaluation is also skewed. If, for example, only data for men is used as the basis for an analysis process, women may be put at a disadvantage.

It is also dangerous if the AI systems have not been taught all the relevant criteria. For instance, the Medical Center at the University of Pittsburgh noted that a major risk factor for severe complications was missing from an AI system for initially assessing pneumonia patients. And there are many other relevant areas in which AI systems are currently in use without having been put through testing and evaluation for bias and inaccuracy.

Checks Needed

As a result of this, the AI Now Institute took to its 2017 research report to call for all important public institutions to immediately stop using “black-box” AI. “When we talk about the risks involved with AI, there is a tendency to focus on the distant future,” says Meredith Whittaker: “But these systems are already being rolled out in critical institutions. We’re truly worried that the examples uncovered so far are just the tip of the iceberg. It’s imperative that we stop using black-box algorithms in core institutions until we have methods for ensuring basic safety and fairness.” 

The post Ethics and principles of AI appeared first on Future Markets Magazine.

Hospitals and doctors’ surgeries have to deal with huge volumes of data: X-ray images, test results, laboratory data, digital patient records, OR reports, and much more. To date, they have mostly been handled separately. But now the trend is towards bringing everything into a single unified software framework. This data integration is not only enabling faster processing of medical data and creating the basis for more efficient interworking between the various disciplines. It is also promising to deliver added value. New, self-learning computing algorithms will be able to detect hidden patterns in the data and provide clinicians with valuable assistance in their diagnostic and therapeutic decision-making.

Better diagnosis thanks to Artificial Intelligence: 30 times faster than a doctor with an error rate of 1%.
Source: PwC

Analysing tissue faster and more accurately

“Artificial Intelligence and robotics offer enormous benefits for our day-to-day work,” asserts Prof. Dr Michael Forsting, Director of the Diagnostic Radiology Clinic of the University Hospital in Essen. The clinic has used a self-learning algorithm to train a system in lung fibrosis. After just a few learning cycles, the computer was making better diagnoses than a doctor: “Artificial Intelligence is helping us to diagnose rare illnesses more effectively, for example. The reasons are that – unlike humans – computers do not forget what they have once learned, and they are better than the human eye at comparing patterns.”

Especially in the processing of image data, cognitive computer assistants are proving helpful in relieving clinicians of protracted, monotonous and recurring tasks, such as accurately tracing the outlines of an organ on a CT scan. The assistants are also capable of filtering information from medical image data that a clinician would struggle to identify on-screen.

Artificial Intelligence diagnosis – Better than the doctor

These systems are now even surpassing humans, as a study at the University of Nijmegen in the Netherlands demonstrates: the researchers assembled two groups to test the detection of cancerous tissue. One comprised 32 -developer teams using dedicated AI software solutions; the other comprised twelve pathologists. The AI developers were provided in advance with 270 CT scans, of which 110 indicated dangerous nodes and 160 showed healthy tissue. These were intended to aid them in training their systems. The result: the best AI system attained virtually 100 per cent detection accuracy and additionally colour-highlighted the critical locations. It was also much faster than a pathologist, who took 30 hours to detect the infected samples with corresponding precision. Most notably, the clinicians overlooked metastases less than 2 millimetres in size under time pressure. Only seven of the 32 AI systems were better than the pathologists, however.

The systems involved in the test are in fact not just research projects, but are already in use. In fibrosis research at the Charité hospital in Berlin, for example, where it is using the Cognitive Workbench from a company called ExB to automate the highly complex analysis of tissue samples for the early detection of pathological changes. The Cognitive Workbench is a proprietary, cloud-based platform which enables users to create and train their own AI-capable analyses of complex unstructured and structured data sources in text and image form. Ramin Assadollahi, CEO and Founder of ExB, states: “In addition to diagnosing hepatic fibrosis, we can bring our high-quality deep-learning processes to bear in the early detection of melanoma and colorectal cancers.”

Cost savings for the healthcare system

According to PwC, AI applications in breast cancer diagnoses mean that mammography results are analysed 30 times faster than by a clinician – with an error rate of just one per cent. There are prospects for huge progress, not only in diagnostics. In a pilot study, Artificial Intelligence was able to predict with greater than 70 per cent accuracy how a patient would respond to two conventional chemotherapy procedures. In view of the prevalence of breast cancer, the PwC survey reports that the use of AI could deliver huge cost savings for the healthcare system. It estimates that over the next 10 years, cumulative savings of EUR 74 billion might be made.

Digital assistants for patients

AI is also benefiting patients in very concrete ways to overcome a range of difficulties in their everyday lives, such as visual impairment, loss of hearing or motor diseases. The “Seeing AI” app, for example, helps the visually impaired to perceive their surroundings. The app recognises objects, people, text or even cash on a photo that the user takes on his or her smartphone. The AI-based algorithm identifies the content of the image and describes it in a sentence which is read out to the user. Other examples include smart devices such as the “Emma Watch”, which intelligently compensates for the tremors typical to Parkinson’s disease patients. Microsoft developer Haiyan Zhang developed the smart watch for graphic designer Emma Lawton, who herself suffers from Parkinson’s. More Parkinson’s patients will be provided with similar models in future.

The post AI better than the doctor? appeared first on Future Markets Magazine.

AI-supported applications must keep pace with rapidly growing data volumes and often have to respond simultaneously in real time. The classic CPUs that you will find in every computer quickly reach their limits in this area because they process tasks sequentially. Significant improvements in performance, particularly in the context of deep learning, would be possible if the individual processes could be executed in parallel.

Hardware for parallel computing processes

A few years, ago, the AI sector focused its attention on the graphics processing unit (GPU), a chip that had actually been developed for an entirely different purpose. It offers a massive parallel architecture, which can perform computing tasks in parallel using many smaller yet still efficient computer units. This is exactly what is required for deep learning. Manufacturers of graphics processing units are now building GPUs specifically for AI applications. A server with just one of these high-performance GPUs has a throughput 40 times greater than that of a dedicated CPU server.

However, even GPUs are now proving too slow for some AI companies. This in turn is having a significant impact on the semiconductor market. Traditional semiconductor manufacturers are now being joined by buyers and users of semiconductors – such as Microsoft, Amazon and even Google – who are themselves becoming semiconductor manufacturers (along with companies who want to produce chips to their own specifications). For example, Alphabet, the parent company behind Google, has developed its own Application-Specific Integrated Circuit (ASIC), which is specifically tailored to the requirements of machine learning. The second generation of this tensor processing unit (TPU) from Alphabet offers 180 teraflops of processing power, while the latest GPU from Nvidia offers 120 teraflops. Flops (Floating Point Operations Per Second) indicate how many simple mathematical calculations, such as addition or multiplication, a computer can perform per second.

Different performance requirements

Flops are not the only benchmark for the processing power of a chip. With AI processors, a distinction is made between performance in the training phase, which requires parallel computing processes, and performance in the application phase, which involves putting what has been learned into practice – known as inference. Here the focus is on deducing new knowledge from an existing database through inference. “In contrast to the massively parallel training component of AI that occurs in the data centre, inferencing is generally a sequential calculation that we believe will be mostly conducted on edge devices such as smartphones and Internet of Things, or IoT, products,” says Abhinav Davuluri, analyst at Morningstar, a leading provider of independent investment research. Unlike cloud computing, edge computing involves decentralised data processing at the “edge” of the network. AI technologies are playing an increasingly important role here, as intelligent edge devices such as robots or autonomous vehicles do not have to transfer data to the cloud before analysis. Instead, they can acquire the data directly on site – saving the time and energy required for transferring data to the data centre and back again.

Solutions for edge computing

For these edge computing applications, another new chip variant – Field-Programmable Gate Array (FPGA) – is currently establishing itself alongside CPUs, GPUs and ASICs. This is an integrated circuit, into which a logical circuit can be loaded after manufacturing. Unlike processors, FPGAs are truly parallel in nature thanks to their multiple programmable logic blocks, which mean that different processing operations are not assigned to the same resource. Each individual processing task is assigned to a dedicated area on a chip and can thus be performed autonomously. Although they do not quite match the processing power of a GPU in the training process, they rank higher than graphics processing units when it comes to inference. Above all, they consume less energy than GPUs, which is particularly important for applications on small, mobile devices. Tests have shown that FPGAs can detect more frames per second and watt than GPUs or CPUs, for example. “We think FPGAs offer the most promise for inference, as they can be upgraded while in the field and could provide low latencies if located at the edge alongside a CPU,” says Morningstar analyst Davuluri.

More start-ups are developing Artificial Intelligence chips

More and more company founders – and investors – are recognising the opportunities offered by AI chips. At least 45 start-ups are currently working on corresponding semiconductor solutions, while at least five of these have received more than USD 100 million from investors. According to market researchers at CB Insights, venture capitalists invested more than USD 1.5 billion in chip start-ups in 2017 – double the amount that was invested just two years ago. British firm Graphcore has developed the Intelligence Processing Unit (IPU), a new technology for accelerating machine learning and Artificial Intelligence (AI) applications. The AI platform of US company Mythic performs hybrid digital/analogue calculations in flash arrays. The inference phase can therefore take place directly within the memory, where the “knowledge” of the neural network is stored, offering benefits in terms of performance and accuracy. China is one of the most active countries when it comes to Artificial Intelligence chip start-ups. The value of Cambricon Technologies alone is currently estimated at USD 1 billion. The start-up has developed a neural network processor chip for smartphones, for instance.

New chip architectures for even better performance of Artificial Intelligence

Neuromorphic chips are emerging as the next phase in chip development. Their architecture mimics the way the human brain works in terms of learning and comprehension. A key feature of these chips is the removal of the separation between the processor unit and the data memory. Launched in 2017, neuromorphic test chips with over 100,000 neurons and 100 million plus synapses can unite training and inference on one chip. When in use, they should be able to learn autonomously at a rate that is a 1 million times better than the third generation of neural networks. At the same time, they are highly energy-efficient.

Quantum computers represent a quantum leap for AI systems in the truest sense of the word. The big players in the IT sector, such as Google, IBM and Microsoft, as well as countries, intelligence services and even car manufacturers are investing in this technology. These computers are based on the principles of quantum mechanics. A quantum computer can perform each calculation step for all states at the same time. This means that it delivers exceptional processing power for the parallel processing of commands and has the potential to compute at a much higher speed than conventional computers. Although the technology may still be in its infancy, the race for faster and more reliable quantum processors is already well underway.

The post Chips driving Artificial Intelligence appeared first on Future Markets Magazine.

The retail sector involves a complex interaction between customers, manufacturers, logistics service providers and online platforms. To gain a competitive edge, retailers need to gauge their customers’ needs optimally, and fulfil them as efficiently and closely as possible. That means retailers have to make the right choices to find the ideal mix of partners. Self-learning algorithms and AI are opening up new dimensions in process optimisation, personalisation and decision-making accuracy.

Artificial Intelligence enables retailers to respond better to their customers’ needs and, for example, optimise their ordering and delivery processes.

Artificial Intelligence is a question of necessity

Prof. Dr Michael Feindt, founder of Blue Yonder: “Anyone who fails to adopt AI will die! But those who open themselves up to the new technology and make smart use of it will have every chance of achieving sustained success in the retail sector. For retailers, digital transformation through AI is not a question of choice, but of necessity. Only those who change and adopt the new AI technologies will survive.” One way that Blue Yonder is responding to that necessity is with a machine learning solution which optimises sales throughout the season based on automated pricing and discounting. The system measures the correlation between price changes and demand trends at each physical outlet and through each channel. Based on the results, the solution automatically sets prices to increase turnover or profit throughout the selling cycle, including the application of discounted pricing and running sale campaigns as appropriate. It analyses both historical and current sales and product master data, and enables hundreds of prices to be validated and optimised each day. Using such systems, retailers can meet consumers’ rising expectations and maximise their profits at the same time. According to Blue Yonder, this means profit can be improved by 6 per cent, sales turnover increased by 15 per cent, and stocks cut by 15 per cent.

Optimising processes in retail with AI

“Artificial Intelligence enables retailers to respond better to their customers’ needs and, for example, optimise their ordering and delivery processes,” says Stephan Tromp, Chief Executive Director of HDE, the German Retail Association. For example, retailers can use their suppliers’ data to measure performance and optimise processes. Combined with the data from their outlets and warehouses, they can also balance supply and demand more closely. For instance, intelligent forecasting systems learn from past orders, create buyer groups, and analyse seasonal effects. From their findings, they can forecast product sales volumes and ideally know before the consumer what he or she is going to order next. This means retailers can tailor their websites to the relevant product groups, trigger purchasing, top up stocks accordingly, and ultimately cut shipping lead times. As a result, bottlenecks in the supply of specific products can be predicted, and retailers can quickly identify which supplier is currently able to deliver top-up stocks of the required merchandise most quickly.

Keeping track of customers’ movements

AI not only has applications in retailers’ back-office operations, however. In the physical shops, too, deep-learning functions are helping to gauge customers’ behaviour. A company called Retailnext, for example, has launched an all-in-one IoT sensor which monitors customers’ movements when in the outlet: collecting their goods, trying on clothing, and walking around the shop. All those movements are monitored by a camera, and analysed directly in the unit with the aid of deep-learning functions. The data is then uploaded to the cloud in real time, so companies can gather valuable information on all the branches in their chain. “It’s precisely those projects that enable retailers to develop a deeper understanding of in-store shopping behaviours and allow them to produce differentiated in-store shopping experiences,” asserts Arun Nair, Co-Founder and Technical Director of Retailnext. “The more retailers know about what’s happening in store, the better.”

The post Retail: “Anyone who fails to adopt AI will die!” appeared first on Future Markets Magazine.